Repo Overview

Preprocessing Scripts

• convert_csv_to_mlflow.py - This script converts Label-Studio Annotations into csv
• convert_pascal_to_googlecsv.py - Powerful script for converting the csv into expected format, dataset splitting and class merging
• png_to_jpeg.py - Bulk convertion from png images to jpegs images
• preproc_imgs.py - Shrinks images to a max width while keeping aspect ratio
• normalize_csv.py - Normalizes bbox coordinates
• remove_other.py - Removes label occurrences of desired class from the dataset
• train_and_run_od_model.ipynb Recommended for proprocessing: Performs all preproc steps and starts training

Train scripts

• do_inference.py - Uses CPU Delegate to perform inference using a tflite model and the label list
• train_and_run_od_model.ipynb Includes proprocessing: Performs all preproc steps and starts training

Cross Validation scripts

• cross_validation_fold_creator.py - Splits a regular dataset to balanced (K-Stratfolds) Cross validation datasets ready for training
• cross_valid_script.py - Uses the created CV datasets to train, performas eval and saves the tflite model

Augmentation & Annotation Inspection helper scripts

• img_augmentation.py - Creates Augmented copies for a CV dataset for the train folds only
• inference_ui.py - Can be used to compare inference results from different model versions
• orig_anno_viewer.py - Used to inspect annotations in the label studio format
• augment_viewer.py - Used to inspect annotations of shrinked images vs their augmented copies
• bbox_draw_test.py - Augmentation test script for a single image and fixed augmentation seed

Legacy scripts (might not work as expected)

• train_old.py
• train_cross_validation.ipynb
• do_inference_other.py

tflite_model_maker_wsl2

This shows how to set up tflite model maker on WSL2

Training Baseline

• Create and label dataset with Label Studio (Needed to use Firefox bc of Request Stalling)
• Download dataset in PascalVoc Format
• Actually better is to load in csv using

    curl -X GET http://url:8080/api/projects/1/export?exportType=CSV -H 'Authorization: Token xxx' --output 'annotations.csv' 
  

  and then convert it with

    import pandas as pd
    
    # Read the provided CSV file
    df = pd.read_csv("your_csv_file.csv")
    
    # Initialize a list to store the converted annotations
    mlflow_annotations = []
    
    # Iterate through each row of the DataFrame
    for _, row in df.iterrows():
        image_path = row["image"]
        annotations = json.loads(row["label"])  # Assuming the label column contains JSON-encoded data
    
        for annotation in annotations:
            label = annotation["rectanglelabels"][0]  # Assuming only one label per annotation
            x = annotation["x"]
            y = annotation["y"]
            width = annotation["width"]
            height = annotation["height"]
            
            # Calculate x_max and y_max based on x, y, width, and height
            x_max = x + width
            y_max = y + height
            
            mlflow_annotations.append([image_path, label, x, y, "", "", x_max, y_max, "", ""])
    
    # Create a DataFrame from the converted annotations
    mlflow_df = pd.DataFrame(mlflow_annotations, columns=["path", "label", "x_min", "y_min", "", "", "x_max", "y_max", "", ""])
    
    # Save the DataFrame to a CSV file in the desired format
    mlflow_df.to_csv("mlflow_annotations.csv", index=False)

  Then download the images.
  • First compress via tar -czvf /home/alex/all_imgs_11_10_23.tar.gz .
  • Then download with scp alex@foodsnapai.com:/home/alex/all_imgs_11_10_23.tar.gz allImgs.tar.gz
  • Then extract with tar -xzvf allImgs.tar.gz -C allImgs_extracted/
• Extract archive then convert to MLFlow format and perform dataset split

> python .\convert_pascal_to_googlecsv.py project-1-at-2023-05-16-11-39-d4943046\Annotations\
Merged CSV file saved as merged_annotations_mlflow.csv
Create train-test-validation split? (y/n): y
Enter the proportion for train set (0 to 1): 0.8
Enter the proportion for test set (0 to 1): 0.1
Enter the proportion for validation set (0 to 1): 0.1
Train-test-validation split applied to merged_annotations_mlflow.csv

• Convert all pngs to jpegs

> python png_to_jpeg.py merged_annotations_mlflow.csv

• Optionally check your dataset stats

> python get_label_stats.py merged_annotations_mlflow.csv

• If you havent already open WSL
• Activate the env to your model maker installation

conda activate conda_env

• Load the pretrained model

import numpy as np
import os

from tflite_model_maker.config import QuantizationConfig
from tflite_model_maker.config import ExportFormat
from tflite_model_maker import model_spec
from tflite_model_maker import object_detector

import tensorflow as tf
assert tf.__version__.startswith('2')

tf.get_logger().setLevel('ERROR')
from absl import logging
logging.set_verbosity(logging.ERROR)

spec = model_spec.get('efficientdet_lite0')

• Load the dataset

train_data, validation_data, test_data = object_detector.DataLoader.from_csv('merged_annotations_mlflow.csv')

• Train the model

model = object_detector.create(train_data, model_spec=spec, batch_size=8, train_whole_model=True, validation_data=validation_data)

• Run on test data

model.evaluate(test_data)

• Export the model to tflite model

model.export(export_dir='new_model')

• Add the desired classed to inference script

label_map = {
    1: 'rice',2: 'carrot',3: 'strawberry',4: 'potato',
    5: 'grape',6: 'kidney bean',7: 'butter',8: 'water melon',
    9: 'tofu',10: 'lentil',11: 'sweet potato',12: 'chickpea',
    13: 'cherry',14: 'chilli',15: 'avocado',16: 'raspberry',
    17: 'zucchini',18: 'pear',19: 'brocoli',20: 'tomato',
    21: 'mango',22: 'onion',23: 'garlic',24: 'apple',
    25: 'coucous',26: 'quinoa',27: 'cucumber',28: 'lemon',
    29: 'ananas',30: 'plum',31: 'cantaloupe',32: 'califlower',
    33: 'kiwi',34: 'black bean',35: 'green bean',36: 'bell pepper',
    37: 'banana',38: 'spinach',39: 'blackberry',40: 'blueberry',
    41: 'orange',42: 'mushroom'
}
num_classes = 42

• Run inference script for visual output

python do_inference.py --input_img project-1-at-2023-05-16-11-39-d4943046/images/f2df3e66-00321.jpeg\
                       --model_url /home/alex/new_model/model.tflite

• Update: The inference script accepts now the previously generated csv file as param and performs inference on all images on the test set!

python do_inference.py --input_csv /mnt/z/IdeaRepos/tflite_model_maker_wsl2/merged_annotations_mlflow.csv\  
 --model_url model_c49_180523/model.tflite --output_dir /home/alex/predictions1

Issues:

Currently the inference does not work on the tflite model!:

python /mnt/z/IdeaRepos/tflite_model_maker_wsl2/do_inference_other.py --input_csv /mnt/c/Users/Fistus/annotations_mlflow_shuffled_n.csv --output_dir /home/alex/test_out --model_url /home/alex/model_ed2_900_e50
/home/alex/anaconda3/envs/conda_env/lib/python3.8/site-packages/tensorflow_addons/utils/ensure_tf_install.py:53: UserWarning: Tensorflow Addons supports using Python ops for all Tensorflow versions above or equal to 2.9.0 and strictly below 2.12.0 (nightly versions are not supported).
 The versions of TensorFlow you are currently using is 2.8.4 and is not supported.
Some things might work, some things might not.
If you were to encounter a bug, do not file an issue.
If you want to make sure you're using a tested and supported configuration, either change the TensorFlow version or the TensorFlow Addons's version.
You can find the compatibility matrix in TensorFlow Addon's readme:
https://github.com/tensorflow/addons
  warnings.warn(
Traceback (most recent call last):
  File "/mnt/z/IdeaRepos/tflite_model_maker_wsl2/do_inference_other.py", line 190, in <module>
    main(args)
  File "/mnt/z/IdeaRepos/tflite_model_maker_wsl2/do_inference_other.py", line 150, in main
    interpreter = tf.lite.Interpreter(model_path=model_path)
  File "/home/alex/anaconda3/envs/conda_env/lib/python3.8/site-packages/tensorflow/lite/python/interpreter.py", line 456, in __init__
    _interpreter_wrapper.CreateWrapperFromFile(
ValueError: Mmap of '3' at offset '0' failed with error '19'.

Already on export something could be wrong:

>>> model.export(export_dir='model_ed2_900_e50')
2023-11-11 02:36:17.027448: W tensorflow/python/util/util.cc:368] Sets are not currently considered sequences, but this may change in the future, so consider avoiding using them.
2023-11-11 02:36:34.875824: W tensorflow/core/common_runtime/graph_constructor.cc:803] Node 'resample_p7/PartitionedCall' has 1 outputs but the _output_shapes attribute specifies shapes for 3 outputs. Output shapes may be inaccurate.

Train with checkpoints

In order to run with checkpoints and check training progress with tensorboard do this

python train_od.py

In a new command line open tensorboard at the checkpoint directory

conda activate conda_env
tensorboard serve --logdir .

Now you should be able to visualize train and val loss during training :)

Installation of TF+Cuda+Cudnn (Way 1)

An alternative as the one from pip is to use Conda. There are tf packages with bundles cudnn and cuda and it even worked on an AVX GPU!

conda activate tf_gpu_env
(tf_gpu_env) alex@DESKTOP-IMOJ0PL:~$ conda install tensorflow=2.8.2=gpu_py38h75b8afa_0

AND IT WORKED:

>>> import tensorflow as tf
>>> tf.config.list_physical_devices('GPU')
2023-12-18 01:18:17.785271: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:922] could not open file to read NUMA node: /sys/bus/pci/devices/0000:02:00.0/numa_node
Your kernel may have been built without NUMA support.
2023-12-18 01:18:17.799686: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:922] could not open file to read NUMA node: /sys/bus/pci/devices/0000:02:00.0/numa_node
Your kernel may have been built without NUMA support.
2023-12-18 01:18:17.800511: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:922] could not open file to read NUMA node: /sys/bus/pci/devices/0000:02:00.0/numa_node
Your kernel may have been built without NUMA support.
[PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]

This is the easiest installation of Cuda+CudNN from all! However I tested it after installing Cuda 11.2 and Cudnn(archive) manually and setting paths. But since the Versions are a bit different I wonder if that was even necessary. If I can reproduce this again on a minimal WSL this would be by far the easiest installation

Installation of model maker 0.4.3 after installation of cuda tensorflow (Way 1)

Installation (Way 2)

0.) You need python38 for doing this! Oh and dont even think of building python3810 from source! It will fuck everything up!

better: create conda env with python3.8.10

You can list your conda envs with the command conda env list

For GPU Support you need WSL2 and Kernel version higher than 5.10.16.3-microsoft-standard-WSL2

Set the default version of WSL to WSL2 by doing this:

wsl --set-default-version 2

Then install ubuntu.

Tried with

sudo cat /proc/version
Linux version 5.15.79.1-microsoft-standard-WSL2 

and now nvidia-smi works!

1.) I created a conda env conda_env and installed tensorflow==2.8.4 and packaging==20.9 inside it

1.1) then I cloned the repo and installed it

git clone https://github.com/tensorflow/examples
cd examples/tensorflow_examples/lite/model_maker/pip_package
pip install -e .

2.) It seems to be installed now

pip list
tflite-model-maker            0.4.2     /home/alex/examples/tensorflow_examples/lite/model_maker/pip_package/src

2.1) I got an error importing something in modelmaker so I downgraded numpy to 1.23 Then I continued with this

2.2) Now I got the following error which I am unable to fix

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/alex/anaconda3/envs/just_mm/lib/python3.8/site-packages/tflite_model_maker/__init__.py", line 51, in <module>
    from tflite_model_maker import searcher
  File "/home/alex/anaconda3/envs/just_mm/lib/python3.8/site-packages/tflite_model_maker/searcher/__init__.py", line 25, in <module>
    from tensorflow_examples.lite.model_maker.core.task.searcher import ExportFormat
  File "/home/alex/anaconda3/envs/just_mm/lib/python3.8/site-packages/tensorflow_examples/lite/model_maker/core/task/searcher.py", line 30, in <module>
    from tensorflow_examples.lite.model_maker.core.utils import ondevice_scann_builder
  File "/home/alex/anaconda3/envs/just_mm/lib/python3.8/site-packages/tensorflow_examples/lite/model_maker/core/utils/ondevice_scann_builder.py", line 17, in <module>
    from scann.proto import scann_pb2
  File "/home/alex/anaconda3/envs/just_mm/lib/python3.8/site-packages/scann/__init__.py", line 2, in <module>
    from scann.scann_ops.py import scann_ops
  File "/home/alex/anaconda3/envs/just_mm/lib/python3.8/site-packages/scann/scann_ops/py/scann_ops.py", line 23, in <module>
    _scann_ops_so = tf.load_op_library(
  File "/home/alex/anaconda3/envs/just_mm/lib/python3.8/site-packages/tensorflow/python/framework/load_library.py", line 54, in load_op_library
    lib_handle = py_tf.TF_LoadLibrary(library_filename)
tensorflow.python.framework.errors_impl.NotFoundError: /home/alex/anaconda3/envs/just_mm/lib/python3.8/site-packages/scann/scann_ops/cc/_scann_ops.so: undefined symbol: _ZN4absl12lts_2021032420raw_logging_internal21internal_log_functionE

import numpy as np
import os

from tflite_model_maker.config import QuantizationConfig
from tflite_model_maker.config import ExportFormat
from tflite_model_maker import model_spec
from tflite_model_maker import object_detector

import tensorflow as tf
assert tf.__version__.startswith('2')

tf.get_logger().setLevel('ERROR')
from absl import logging
logging.set_verbosity(logging.ERROR)

Probably because I used the conda tensorflow but honestly I dont know. this also happens when installing model maker from pip.

spec = model_spec.get('efficientdet_lite0')

3.) The following results in an error making be think it could not load the data from the gcs bucket

train_data, validation_data, test_data = object_detector.DataLoader.from_csv('gs://cloud-ml-data/img/openimage/csv/salads_ml_use.csv')

Or if you have already downloaded the images using the python script use just your generated csv and save time

train_data, validation_data, test_data = object_detector.DataLoader.from_csv('output.csv')

4.) But then running this seems to train correctly (currently on cpu only I think, cuda11 could not be found)

model = object_detector.create(train_data, model_spec=spec, batch_size=8, train_whole_model=True, validation_data=validation_data)

Btw: Here on a fresh installed ubuntu22 with WSL2 I got the following error UnicodeEncodeError: 'utf-8' codec can't encode character '\udcfc' in position 5: surrogates not allowed

Solution: This was because the hostname on Windows which contained non-ASCII characters was translated wrongly to Ubuntu with an invalid character.

Running sudo hostname NewHostName and restarting wsl fixed it!

To persist this create a startup script

sudo nano /etc/init.d/hostname

With these lines

#!/bin/sh
sudo hostname NewHostName

Make it executable

sudo chmod +x /etc/init.d/hostname

Update the startup script by running the following command:

sudo update-rc.d hostname defaults

It trained fine for a while but then failed because of

ImportError: You must install pycocotools (`pip install pycocotools`) (see github repo at https://github.com/cocodataset/cocoapi) for efficientdet/coco_metric to work.

4.1) so I installed it and also installed the next required module:

pip install -q pycocotools
pip install -q opencv-python-headless==4.1.2.30

In case pycocotools installation throws an error you might need to install these

sudo apt-get install libpq-dev python3.8-dev libxml2-dev libxslt1-dev libldap2-dev libsasl2-dev libffi-dev libjpeg-dev zlib1g-dev
sudo apt-get install build-essential

then I run the code again and restarted training

spec = model_spec.get('efficientdet_lite0')

It worked!

5.) Then I ran eval

model.evaluate(test_data)

6.) worked also and then I exported the model

model.export(export_dir='.')

Worked and resulted in:

{'AP': 0.20119594, 'AP50': 0.34621373, 'AP75': 0.21721725, 'APs': -1.0, 'APm': 0.5330238, 'APl': 0.19910036, 'ARmax1': 0.17052028, 'ARmax10': 0.34225824, 'ARmax100': 0.39623663, 'ARs': -1.0, 'ARm': 0.7, 'ARl': 0.39222324, 'AP_/Baked Goods': 0.025994863, 'AP_/Salad': 0.5475503, 'AP_/Cheese': 0.19085687, 'AP_/Seafood': 0.021515297, 'AP_/Tomato': 0.22006239}

7.) Then I loaded the just exported tflite model to perform evaluation:

model.evaluate_tflite('model.tflite', test_data)

worked as well!

{'AP': 0.1789487, 'AP50': 0.31319344, 'AP75': 0.1972646, 'APs': -1.0, 'APm': 0.559314, 'APl': 0.17661628, 'ARmax1': 0.13116644, 'ARmax10': 0.24777183, 'ARmax100': 0.26256573, 'ARs': -1.0, 'ARm': 0.64166665, 'ARl': 0.25821817, 'AP_/Baked Goods': 0.0, 'AP_/Salad': 0.5157693, 'AP_/Cheese': 0.17967021, 'AP_/Seafood': 0.0006441821, 'AP_/Tomato': 0.19865985}

Train on your GPU in WSL

You need WSL2 and Kernel version higher than 5.10.16.3-microsoft-standard-WSL2

Installing cuda on wsl2

I need tensorflow 2.8 therefor Cuda 11.2 from the Nvidia page

sudo wget https://developer.download.nvidia.com/compute/cuda/repos/wsl-ubuntu/x86_64/cuda-wsl-ubuntu.pin
sudo mv cuda-wsl-ubuntu.pin /etc/apt/preferences.d/cuda-repository-pin-600
sudo wget https://developer.download.nvidia.com/compute/cuda/11.2.0/local_installers/cuda-repo-wsl-ubuntu-11-2-local_11.2.0-1_amd64.deb
sudo dpkg -i cuda-repo-wsl-ubuntu-11-2-local_11.2.0-1_amd64.deb
sudo apt-key add /var/cuda-repo-wsl-ubuntu-11-2-local/7fa2af80.pub
sudo apt-get update
sudo apt-get -y install cuda

then install cudnn

Important: Use the archieve file and extract it Use cudnn 8.6 then try it out and it should show this:

import tensorflow as tf
tf.config.list_physical_devices('GPU')
2023-02-07 11:27:39.769247: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:922] could not open file to read NUMA node: /sys/bus/pci/devices/0000:09:00.0/numa_node
Your kernel may have been built without NUMA support.
2023-02-07 11:27:39.774772: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:922] could not open file to read NUMA node: /sys/bus/pci/devices/0000:09:00.0/numa_node
Your kernel may have been built without NUMA support.
2023-02-07 11:27:39.774827: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:922] could not open file to read NUMA node: /sys/bus/pci/devices/0000:09:00.0/numa_node
Your kernel may have been built without NUMA support.
[PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]

current error: when trying to train i get an error even though the file mentioned can be found in the lib64 dir

>>> model = object_detector.create(train_data, model_spec=spec, batch_size=8, train_whole_model=True, validation_data=validation_data)
Epoch 1/50
2023-02-07 11:33:29.041737: W tensorflow/core/framework/cpu_allocator_impl.cc:82] Allocation of 23970816 exceeds 10% of free system memory.
2023-02-07 11:33:29.049214: W tensorflow/core/framework/cpu_allocator_impl.cc:82] Allocation of 26873856 exceeds 10% of free system memory.
2023-02-07 11:33:29.054961: W tensorflow/core/framework/cpu_allocator_impl.cc:82] Allocation of 36499968 exceeds 10% of free system memory.
2023-02-07 11:33:29.055949: W tensorflow/core/framework/cpu_allocator_impl.cc:82] Allocation of 24443136 exceeds 10% of free system memory.
2023-02-07 11:33:29.066611: W tensorflow/core/framework/cpu_allocator_impl.cc:82] Allocation of 51556935 exceeds 10% of free system memory.
2023-02-07 11:33:32.566107: I tensorflow/stream_executor/cuda/cuda_dnn.cc:368] Loaded cuDNN version 8600
Could not load library libcudnn_cnn_infer.so.8. Error: libcuda.so: cannot open shared object file: No such file or directory
Aborted

Solution: This is WSL related! Do this:

export LD_LIBRARY_PATH=/usr/lib/wsl/lib:$LD_LIBRARY_PATH

Be sure that your library is in /usr/lib/wsl/lib, to see it you can run

ldconfig -p | grep cuda

In the end the environment vars in .bashrc should look like this

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda-11.2/lib64:/usr/lib/wsl/lib
export CUDA_HOME=/usr/local/cuda-11.2
export PATH=$PATH:$CUDA_HOME/bin

Now the training should work! :)

Create own dataset

1. Load and install Label Studio

pip install label-studio
label-studio start

Then go to http://localhost:8080/, create a project, upload data and label it

2. Export the labels to Pascal VOC XML

3. Then use the script convert_pascal_to_googlecsv.py to convert the xmls into the Google OD csv format.

4. Assign labels for TRAIN, TEST, VAL

5. Dont forget to check your dataset labels. Each label should have more than 100 occurences in test images! You can use the get_label_stats.py script in this repo to check this.

6. Start training on your dataset! :)

Tried on windows (legacy)

(model_maker_venv) (base) PS Z:\IdeaRepos\model_maker_od> pip install -q --use-deprecated=legacy-resolver tflite-model-maker
ERROR: Could not find a version that satisfies the requirement scann==1.2.6 (from tflite-model-maker) (from versions: none)
ERROR: No matching distribution found for scann==1.2.6 (from tflite-model-maker)
WARNING: You are using pip version 22.0.4; however, version 23.0 is available.

You should consider upgrading via the 'Z:\IdeaRepos\model_maker_od\model_maker_venv\Scripts\python.exe -m pip install --upgrade pip' command.

--> scann requires linux -> Do this on WSL2

NEXT STEP (legacy):

Do all of this model maker repo in WSL

sudo apt -y install libportaudio2
pip install -q --use-deprecated=legacy-resolver tflite-model-maker
pip install -q pycocotools
pip install -q opencv-python-headless==4.1.2.30
pip uninstall -y tensorflow && pip install -q tensorflow==2.8.0

Error at installing tflite-model-maker

Tried

pip install tflite-model-maker-nightly

with tensorflow2.8 installed in venv

python3 -m venv model_maker_venv
source model_maker_venv/bin/activate

Two dependency problems

ERROR: grpcio-status 1.51.1 has requirement protobuf>=4.21.6, but you'll have protobuf 3.20.3 which is incompatible.
ERROR: tensorflowjs 3.18.0 has requirement packaging~=20.9, but you'll have packaging 23.0 which is incompatible.

Installed packaging - no error

pip install packaging==20.9

Installed protobuf...

pip install protobuf==4.21.6

with errors:

ERROR: tflite-support-nightly 0.4.4.dev20221103 has requirement protobuf<4,>=3.18.0, but you'll have protobuf 4.21.6 which is incompatible.
ERROR: tensorflow-metadata 1.12.0 has requirement protobuf<4,>=3.13, but you'll have protobuf 4.21.6 which is incompatible.

Updated pip and created new venv with just this

$ pip install tflite-model-maker

resulted in an error

ERROR: tensorflow 2.11.0 has requirement protobuf<3.20,>=3.9.2, but you'll have protobuf 3.20.3 which is incompatible.
ERROR: scann 1.2.6 has requirement tensorflow~=2.8.0, but you'll have tensorflow 2.11.0 which is incompatible.
ERROR: tensorflowjs 3.18.0 has requirement packaging~=20.9, but you'll have packaging 23.0 which is incompatible.

Another aproach (legacy)

Permission errors seem to be WSL related because I was in NTFS file system managed by windows Tried this inside (WSL) /home/alex inside a venv (model_maker_venv)

git clone https://github.com/tensorflow/examples
cd examples/tensorflow_examples/lite/model_maker/pip_package
pip install -e .

ERROR: launchpadlib 1.10.13 requires testresources, which is not installed.
ERROR: tensorboard 2.11.2 has requirement protobuf<4,>=3.9.2, but you'll have protobuf 4.21.12 which is incompatible.
ERROR: tensorflow 2.11.0 has requirement protobuf<3.20,>=3.9.2, but you'll have protobuf 4.21.12 which is incompatible.
ERROR: scann 1.2.6 has requirement tensorflow~=2.8.0, but you'll have tensorflow 2.11.0 which is incompatible.
ERROR: tensorflow-metadata 1.12.0 has requirement protobuf<4,>=3.13, but you'll have protobuf 4.21.12 which is incompatible.
ERROR: tensorflowjs 3.18.0 has requirement packaging~=20.9, but you'll have packaging 23.0 which is incompatible.
ERROR: tflite-support 0.4.3 has requirement protobuf<4,>=3.18.0, but you'll have protobuf 4.21.12 which is incompatible.

Then did

pip install tensorflow~=2.8.0 #(this fixed protobuf it seemes)
pip install packaging~=20.9
pip install -e .
Successfully installed tflite-model-maker

Then continued model-maker installation with

pip install -q pycocotools
pip install -q opencv-python-headless==4.1.2.30

Skipped this because I think i did it already by installing tensorflow-2.8 before

pip uninstall -y tensorflow && pip install -q tensorflow==2.8.0 # didnt do this

This way seems to install model maker locally so I switched into the src dir and invoked python3 and tried to import model-maker but I resulted in error that tensorflowjs was missing so maybe this wasnt installed correctly.